Fisheye Image Object Detection Based on an Improved YOLOv3 Algorithm

摘要

The accuracy and speed of object detection based on deep learning are much higher than that of human eyes, but the application of deep learning in object detection of fisheye image remains to be studied. Although the fisheye image has a wide field of vision, it has the problem of geometric distortion. In order to solve the problem that it is difficult to obtain features caused by fisheye image distortion, the 8-fold and 16-fold down-sampling feature images obtained by YOLOv3 for fisheye image detection were respective Maxpool to the same size as the 32-fold down-sampling feature images for fusion, so as to realize feature reuse and increase the detail features in the 32-fold down-sampling feature images. This paper constructs its own fisheye image data set and carries out experiments. The experimental results show the improved YOLOv3 algorithm has higher object detection accuracy for fisheye images in comparison to the original algorithm.